Introduction
Methane ($CH_4$) is a critical target for climate change mitigation, possessing a global warming potential 85 times greater than carbon dioxide over a 20-year period. While satellite instruments like the Sentinel-5P TROPOMI provide daily global coverage of methane concentrations, the resulting data is often sparse and highly susceptible to atmospheric noise. A significant hurdle in utilizing this data for source attribution is atmospheric transport; wind and diffusion "smear" methane plumes away from their physical sources, making simple temporal averages ineffective for pinpointing leak locations.
The Challenge of Atmospheric Transport
Traditional methods for aggregating satellite imagery rely on arithmetic (pixel-wise) averaging. However, when gas is transported by wind, an arithmetic mean results in a diffuse cloud that fails to represent the actual emission source. Furthermore, satellite observations are frequently interrupted by cloud cover or water surfaces, leading to missing data. Standard conservation-of-mass models struggle with these discontinuities and the inherent variability of emission rates.
Methodology: Optimal Transport and Wasserstein Barycenters
To address these issues, researchers have applied Optimal Transport (OT) theory, specifically Wasserstein barycenters (WB), to average gas concentration maps. Unlike arithmetic means, which treat pixels independently, Wasserstein barycenters consider the underlying geometry of the distribution.
Wind-Aware Cost Functions
The most significant innovation in this work is the integration of historical weather data (ECMWF-ERA5) into the OT cost matrix. By making it "cheaper" to move mass with the wind and "more expensive" to move it against the wind, the barycenter calculation effectively performs an "inverse transport." This pulls the observed methane mass back toward its likely origin.
Unbalanced Optimal Transport
Because methane is constantly being emitted and diffused, and because satellite data has gaps, the total "mass" of gas is not conserved between images. The authors utilize Unbalanced Optimal Transport and the Wasserstein-Fisher-Rao (WFR) metric. This allows for the creation and destruction of mass within the model, accommodating new emissions and missing pixels without breaking the mathematical framework.
Key Results and Case Studies
The effectiveness of the wind-aware WB method was demonstrated across several major methane-emitting regions.
Ohio-Pennsylvania Coal Mines
In the Appalachian region, where coal mines are a major source of methane, prevailing winds typically shift plumes eastward. A standard arithmetic average shows a broad plume centered east of the actual mines. In contrast, the wind-aware Wasserstein average successfully shifted the detected mass westward, aligning it precisely with known mine locations.
Iraq-Kuwait and the Permian Basin
The method was also applied to the Rumaila oil field in the Iraq-Kuwait border region. While classical averaging produced a diffuse signal across the desert, the Wasserstein approach concentrated the methane mass around the specific infrastructure of the field. Similarly, in the Permian Basin, the WFR metric allowed researchers to distinguish between the Delaware and Midland sub-basins despite sparse observations.
The Inverse Transport Shortcut
A key insight of this research is that adapting the cost function of a barycenter calculation can achieve results similar to full atmospheric inversion models. Full inversions are computationally grueling and data-intensive; the WB approach offers a significantly faster alternative for source attribution and infrastructure monitoring.
Limitations and Future Work
Despite its advantages, the method faces challenges. Calculating barycenters with wind-integrated cost matrices is computationally expensive, limiting its application at very high resolutions. Furthermore, the model's accuracy is tethered to the quality of underlying weather data and does not inherently correct for sensor biases like ground albedo. Future research aims to optimize these algorithms for higher resolutions and more complex atmospheric turbulence models.
Summary updated